利用超音波感測及電腦視覺技術經由簡易學習作自動車導航之研究

Abstract

本研究提出了兩套智慧型自動車系統，其一是以視覺為基礎的自動巡邏車，具有利用跟隨人物同時做學習路徑、與人互動、二維影像分析以及路徑規劃等功能；其二為利用超音波訊號分析做智慧型導航的自動車。我們利用一台具有PTZ攝影機的自動車作為實驗平台，並且利用有線網路與自動車通訊，使其航行於室內環境中。針對一利用格狀圖案為校正目標的舊有影像校正方法，我們提出了一精確的十字中心自動偵測方法，增進了校正的準確性與自動化。我們也建立了一計步器校正技術，用一校正模型來減少自動車所累積的行走機械誤差。接著我們發展一利用動態調整跟隨人物衣服顏色參考值的方法，來適應跟隨人物時環境中的不均勻照度。針對跟隨人物在自動車前的快速消失，我們亦提出一利用攝影機轉動角度，紀錄使用者方向的尋找方法，並使用一增加參考圖素數量的方法，來增進偵測人物朝向的準確率。當使用者要求自動車進行學習時，自動車能夠從影像中自動找出需要被監控的物品。此外，我們利用最小化均方差以及動態選擇門檻值，來規劃跟隨人物時所學習的路徑。我們同時提出一分析串列超音波訊號的方法，使自動車能在狹窄的室內空間中進行導航。最後我們以成功的巡邏、跟隨和導覽實驗證明本系統的完整性與可行性。

Two intelligent autonomous vehicle systems for use in indoor environment applications are proposed. One is a vision-based vehicle system for security patrolling which has the capabilities of learning paths by person following, interaction with humans, analyzing 2D images and conducting path planning. Another is an intelligent vehicle system for navigation by analyzing ultrasonic signals. A vehicle equipped with wired control and a web PTZ camera is used as a test bed. First, a method for improving the practicability of an adopted camera calibration method is proposed, which uses a grid pattern as the calibration target and detects the centers of cross shapes on the pattern in a more precise and automatic manner. Next, a technique of odometer calibration is proposed, which uses an odometer calibration model to reduce the incremental mechanical errors the vehicle suffers. Besides, a technique for person following in an environment with variant illuminations is proposed, which adjusts the reference color value for person following dynamically. To follow a target person who turns fast in front of the vehicle, a technique of using the image information exhibited by the angle of panning of the camera to search the disappearing person is proposed. Also proposed is a technique for enhancing the accuracy of detection of the facing direction of a human by increasing the number of pixels for use in the detection work. For object monitoring, the vehicle is enabled to find out the 2D object in the image automatically when the followed person asks it to learn the object. In addition, a minimum MSE technique for refining dynamically the path learned while following a person is proposed. Finally, a method of vehicle navigation in corridor environments by the use of the ultrasonic signal sequence is proposed. Good experimental results show the flexibility and feasibility of the proposed methods for the applications of security patrolling, person following, and people guiding in indoor environments.